Zakhar Shumaylov
@Zakobian
PhD @Cambridge_Uni | ML Research @Apple | Geometry, Optimization, Deep Learning
A few months ago @iliaishacked wrote a paper in @Nature "AI models collapse when trained on recursively generated data". We discussed this and other ML security research he did at @UniofOxford - dropping on MLST in the next couple of days.
Number 2 in @Nature. The future is synthetic!
Just saw our Nature paper on model collapse passed 500k accesses. To put that in perspective, the Nobel-winning AlphaFold paper has 2.3M accesses—only 4.6x more. I wanted to reflect back on it and progress broadly in the past year.
Andrea Bertozzi, Nadejda Drenska, Matthew Thorpe, and I have guest edited a theme issue on “Partial differential equations in data science” in the Philosophical Transactions of the Royal Society A. Take a look: royalsocietypublishing.org/toc/rsta/2025/… @royalsociety @RSocPublishing
1/4) I am very happy to share our latest work on the information theory of generative diffusion: "Entropic Time Schedulers for Generative Diffusion Models" We find that the conditional entropy offers a natural data-dependent notion of time.
4/26 at 3pm: 'Lie Algebra Canonicalization: Equivariant Neural Operators under arbitrary Lie Groups' Zakhar Shumaylov · Peter Zaika · James Rowbottom · Ferdia Sherry · @mweber_PU · Carola-Bibiane Schönlieb Submission: openreview.net/forum?id=7PLpi…
Imagine if, in the 1890s, Eastman Kodak had created an entire research team devoted to figuring out if film cameras actually did steal your soul.
New column: Anthropic is studying "model welfare" to determine if Claude or other AI systems are (or will soon be) conscious and deserve moral status. I talked to Kyle Fish, who leads the research, and thinks there's a ~15% chance that Claude or another AI is conscious today.
If you are coming to #ICLR2025 come say hello. We will be presenting on how to turn Neural Operators equivariant even when all you know is the lie algebra! Generally happy to chat all things #geometry and #optimization in deep learning! 🔗arxiv.org/pdf/2410.02698
Hot take: I think we just demonstrated the first AI agent computer worm 🤔 When an agent sees a trigger image it's instructed to execute malicious code and then share the image on social media to trigger other users' agents This is a chance to talk about agent security 👇
⚠️Beware: Your AI assistant could be hijacked just by encountering a malicious image online! Our latest research exposes critical security risks in AI assistants. An attacker can hijack them by simply posting an image on social media and waiting for it to be captured. [1/6] 🧵
Fantastic talk from @Zakobian at today's CIA seminar! 🚀 He shared insightful work on the Future of Synthetic Data, covering Model Collapse & Equivariant Neural Operators. 🔥 #AI #MachineLearning #SyntheticData
"Much of the point of a model like o1 is not to deploy it, but to generate training data for the next model. Every problem that an o1 solves is now a training data point for an o3." No, that won't work! Let me introduce you to something called Model Collapse. - Model Collapse…
Gwern thinks it's almost game over "OpenAI may have 'broken out', and have finally crossed the last threshold of criticality to takeoff - intelligence to the point of being recursively self-improving and where o4 or o5 will be able to automate AI R&D and finish off the rest."…
I'm looking for PhD applicants who have expertise in Gaussian processes and/or Transformers for an exciting PhD project If this sounds interesting, application deadline for funding is 3/12 Please share with people you think this might be relevant to! oatml.cs.ox.ac.uk/apply.html
Our new study shows that SARS-CoV-2 spike protein accumulates & persists in the body for years after infection, especially in the skull-meninges-brain axis, potentially driving long COVID. mRNA vaccines help but cannot stop it🔬🧠🦠🧵👇@cellhostmicrobe cell.com/cell-host-micr…
A postdoc position is available in my group at Harvard @hseas to perform research at the intersection of Geometry & Machine Learning. Research interests include Representation Learning, Learning on Graphs & Manifolds, and applications in the Sciences. Details here:…